Uncertainty Estimation for Molecules: Desiderata and Methods

Authors: Tom Wollschläger, Nicholas Gao, Bertrand Charpentier, Mohamed Amine Ketata, Stephan Günnemann

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our extensive experimental evaluation, we test four different UE with three different backbones and two datasets. In out-of-equilibrium detection, we find LNK yielding up to 2.5 and 2.1 times lower errors in terms of AUC-ROC score than dropout or evidential regression-based methods while maintaing high predictive performance.
Researcher Affiliation Academia Tom Wollschl ager 1 Nicholas Gao 1 Bertrand Charpentier 1 Mohamed Amine Ketata 1 Stephan G unnemann 1 1Department of Computer Science & Munich Data Science Institute, Technical University of Munich, Germany.
Pseudocode No The paper does not contain any explicitly labeled pseudocode or algorithm blocks. Methods are described textually.
Open Source Code Yes Find our code at cs.cit.tum.de/daml/uncertainty-for-molecules
Open Datasets Yes Datasets. QM7-X: (Hoja et al., 2021) This dataset covers both equilibrium and non-equilibrium structures. We train on equilibrium structures and non-equilibrium structures are considered OOD data. MD17: (Chmiela et al., 2017) MD17 contains energies and forces for molecular dynamics trajectories of different organic molecules.
Dataset Splits Yes Table 10. Hyperparameters of the datasets used with all models: val set size 4151 1000
Hardware Specification No The paper mentions evaluating runtime on QM7X but does not provide specific details about the hardware used, such as GPU or CPU models, or cloud computing instance types.
Software Dependencies No The paper mentions using PyTorch-Geometric in the footnote of Table 11 for Sch Net, but it does not specify version numbers for PyTorch-Geometric, PyTorch, CUDA, or other key software components, which is necessary for reproducibility.
Experiment Setup Yes Table 8 and Table 9 provide specific hyperparameters and settings used for training models on QM7-X and MD17 datasets, including learning rate, patience, force weighting factor, number of inducing points, warmup steps, decay steps, decay rate, EMA decay, and dropout locations.